Scientists Turn to ChatGPT for Daughter's Health Clues

Hilary Eaton was in the middle of a virtual meeting for work when her phone pinged. Glancing down, her eyes locked on the notification: A new test result had arrived in her daughter’s patient portal.

For most of 5-year-old Olivia’s life, the chubby-cheeked girl with blue eyes and a mischievous grin had been afflicted with confounding symptoms. As an infant, she couldn’t sit up on her own. She never slept through the night. Colds lasted for weeks. When she skinned her knees while playing, the wounds wouldn’t heal.

Countless doctors’ visits had yielded few satisfactory answers. Olivia had been diagnosed with a connective tissue disorder. That explained her early physical challenges — her older sister, Fiona, would later be diagnosed with the same condition — but not why she was sick all the time.

Doctors said Olivia’s frequent bouts of illness were likely because she was isolated from other children during the coronavirus pandemic. But Eaton couldn’t shake the feeling something else was going on. Her “spidey sense,” she said, was constantly tingling.

“I knew she was sick too much, too long,” Eaton said.

She hoped the lab tests, ordered by an immunologist, would offer new information.

So on that January day in 2025, as chatter from the work meeting streamed from her computer, Eaton’s attention was on her phone. She zoomed in on a PDF so she could read the test results.

Several numbers stood out: They measured antibodies, the body’s critical line of defense against foreign invaders. Olivia had far less than a healthy child should.

Holy [expletive], Eaton thought.

She texted her husband as questions ran through her mind: What could cause this? Is it dangerous?

Knowing she wouldn’t be able to reach Olivia’s doctor right away, she thought of another way to get information quickly.

As scientists by training, Eaton and her husband, Matt, both used artificial intelligence in their jobs in the biotech industry. They had more recently started asking chatbots specific health questions at home. Now, Eaton wondered if AI could help if she gave it more information — feeding in Olivia’s symptoms and her lab results.

“I’m just going to shove everything into it,” she recalled deciding.

Eaton opened ChatGPT and started to type.

“With the following test results,” she wrote, “what are possible diagnoses?”

The first piece of the puzzle

Using large language models, or LLMs, for health questions has surged in popularity. A quarter of Americans now report using AI chatbots to diagnose symptoms, according to a Pew Research Center survey from June. A similar share of respondents said they use AI to better understand information from their providers.

Yet AI can hallucinate and make mistakes — sometimes serious ones. In recent years, a cancer patient came to believe his treatment could be shortening his life, based in part on conversations with AI. He refused options recommended by his doctor and eventually died. AI companies have also faced lawsuits alleging that interactions with chatbots ended in people dying by suicide.

Experts say chatbots can be helpful tools for health-related questions, if used properly and with caution. They say patients should also understand the privacy policies of chatbots they plan to use with their medical information.

Platforms are also rapidly improving, and health care providers themselves are increasingly using the technology. Today, about 80 percent of physicians say they use AI for work, according to an American Medical Association survey published in March.

Generative AI has “gone from basically no one using it three years ago to being ubiquitous,” said Dr. Adam Rodman, director of AI programs for the Carl J. Shapiro Institute for Education and Research at Beth Israel Deaconess Medical Center. “It’s the fastest that I have ever seen technology proliferate in medicine.”

Every week, more than 300 million people pepper ChatGPT with health queries, said Ashley Alexander, vice president of Health Products at OpenAI. She added that the company takes data privacy seriously and hundreds of physicians provide feedback on the model.

An OpenAI employee would eventually hear Eaton speak about her family’s experience using ChatGPT, and the company has invited them to film a social media video; as compensation for their time, the company will make a donation to a rare-disease advocacy group. (The Eatons were not paid by OpenAI for participating in this story.)

After uploading Olivia’s lab results on that day early last year in her home office in Winchester, Eaton stared at her phone. ChatGPT responded to her query in a matter of seconds.

It was possible, the chatbot said, that her daughter had a primary immunodeficiency called CVID, an acronym for common variable immunodeficiency. Although low antibodies are a symptom of the manageable condition, CVID would turn out not to be the cause.

A mix of feelings swirled through Eaton. Validation and frustration. But also a sense of control.

Now she had something to go off.

“I could go back to the doctor and say, ‘I’ve done my homework. I’ve looked at a bunch of this stuff. Here’s what I’m thinking. Is that in any way aligned with what you’re thinking?’” Eaton said. “It just felt like I could have some iota of agency or control back, that I could start having a conversation where I’m not just sort of the passenger [on] a roller coaster that is totally off the rails.”

‘Genome detectives’

Hilary and Matt Eaton met about two decades ago as graduate students at Duke University. At the time, Hilary said she was taking a break from dating. For how long? Matt asked. Maybe six months, she replied.

Matt set a calendar reminder. On the six-month anniversary of that conversation, he had a reservation at the nicest restaurant in Durham for the two of them.

Persistence and curiosity have defined the couple’s relationship ever since, said Eaton, who is now 42. She graduated with a PhD in molecular cancer biology and completed a postdoctoral fellowship at Harvard Medical School. Matt, 44, also earned his PhD, did his postdoctoral research at MIT, and is a computational biologist.

The couple hoped for two children and in vitro fertilization resulted in two viable embryos. Their first daughter, Fiona, arrived in June 2017. Olivia followed less than two years later.

Soon, the parents grew concerned about Olivia. As a newborn, she failed her hearing test. At 9 months old, she couldn’t sit up on her own. Tests from their doctors came back inconclusive.

Fiona, who has autism, had also hit milestones later than expected, but “not crazy late,” Eaton said. Doctors tried to reassure the couple about Olivia’s development.

“Don’t compare her to her sibling.”

“Every kid’s on a different trajectory.”

Rationally, the explanations made sense. “It’s easy to sort of justify it away,” Eaton said. “But it just continued.”

By the time Olivia was about a year and a half old, she still couldn’t pull herself up or walk. During a 2020 orthopedics appointment for Olivia, the doctor asked her mother an unusual question: Did she have any “party tricks?” Specifically, was Eaton especially bendy?

“I’m double-jointed everywhere,” Eaton recalled saying.

Doctors suspected Olivia might have a connective tissue disorder passed down from her mother. Common symptoms typically include joint hypermobility and fragile tissues, which can lead to frequent bruising or poor wound healing.

But that still didn’t explain why Olivia was getting sick so often.

The parents kept a calendar where they marked the days she didn’t have symptoms of respiratory illness or skin infection. When they added them up, Eaton said, “it was like seven days out of the year that she was healthy.”

They needed more information. “In many ways, parenthood is the epitome of having to let go of control,” Eaton said. “But as a scientist, you want data because that’s how you make decisions, and feeling helpless is like the worst thing in the world.”

In 2023, the Eatons spent about $4,500 to sequence the entire genomes of all four members of their family. Then they got down to work.

After they put the kids to bed, and continuing late into the night, they used ChatGPT to help them identify genes related to the connective tissue disorder. They also pored over the genome sequencing data looking for potential mutations that could help explain her low antibody count. They saved their research in an online folder they labeled “Genome Detectives.”

“This is us just amateur spelunking, basically, in our daughter’s DNA,” Eaton said.

Since the couple knew chatbots were prone to hallucinations and errors, they took steps to train the model. They explained their backgrounds in science, told the chatbot how they wanted it to interact with them, and instructed it to provide research citations for information. They emphasized that they would review the journal articles themselves.

But even with their scientific training, and the safeguards they put in place, the Eatons weren’t immune to scares.

As they searched for mutations that might explain Olivia’s immunodeficiency, ChatGPT flagged several possible genes. When the parents looked up one of the mutations in medical literature, they were horrified.

“You find that these kids are dead by the time they’re like 7 to 10,” Eaton said. “Olivia at the time was 6, so we freaked out. We spent two hours thinking that she was close to death.”

It turned out, however, that the genome sequencing data itself contained errors. Olivia didn’t have that deadly mutation. The chatbot had been working off bad source material.

The Eatons cleaned up the genome data and learned to approach what they were finding with greater caution. But the experience didn’t deter them.

‘Are you sitting down?’

The Eatons and their doctors were beginning to understand Olivia’s condition was more complex than a single disorder. They didn’t always know which symptoms aligned with a particular diagnosis and the chatbot was helping them connect the dots.

It was difficult, Matt Eaton recalled, but the couple was grateful to have their scientific training and experience using AI. “We were obviously having a tough time just with the realities of the diagnosis,” he said. “But at least we can wrap our heads around it.”

The Eatons’ backgrounds make them ideal users of a large language model, several experts said. And rare diseases can be an excellent use case for the technology, said Arjun Manrai, an associate professor of biomedical informatics at Harvard Medical School.

“These are cases where there’s a lot of information, but oftentimes an elusive missing diagnosis,” said Manrai, who researches AI in health care. “A lot of context can be given to these models to help us think through a second opinion or another differential diagnosis, which can then be taken back to physicians.”

AI is now a fixture in the Eatons’ lives. They rely on ChatGPT for helping them research and prepare for doctor appointments, using the chatbot to come up with questions they should ask.

And they’ve seen results.

Olivia’s low antibody levels had been worrying them ever since Eaton received the alarming test results on her phone last January. The chatbot helped them make sense of the results and offered possible treatments to ask about. One of the suggestions, a monthly IV treatment that pumps her full of healthy antibodies, ended up being what doctors recommended.

But that was still just a treatment — it didn’t explain the root cause of Olivia’s low antibodies. More sleuthing with ChatGPT around Olivia’s other conditions would also help solve that mystery.

Eaton discovered another mutation that could explain why Olivia’s scrapes took so long to heal, and why removing a Band-Aid could take off some of her skin. The finding, Eaton said, prompted doctors to order additional genetic testing, which in turn pointed them to a possible answer.

Olivia had Okur-Chung neurodevelopmental syndrome, a rare genetic disorder that is estimated to affect only 1 in 100,000 people. The severity of symptoms can vary, but those with the disorder typically have some degree of developmental delay and differences in brain function. Symptoms can be treated, such as using physical therapy for motor delays. Doctors believed that the genetic mutation could also be affecting Olivia’s ability to make antibodies.

Olivia started the IV treatment last fall, and her parents have seen a significant boost to her immune system.

“She’s missed three days of school,” Eaton said. “That’s it.”

Dr. Wendy Chung, one of Olivia’s doctors and the co-discoverer of Okur-Chung syndrome, said the key for parents using alternative sources of information, like chatbots, is for it to be a partnership with medical providers.

“The more that each of us can do, powered by whatever tools we can do it [with], the better we get to answers, and the faster we get to the right path,” said Chung, chief of the department of pediatrics at Boston Children’s Hospital.

The couple partnered with doctors again this winter, when they noticed Olivia began displaying unusual motor tics, such as grimacing uncontrollably. In preparation for a medical appointment, Eaton once again fired up ChatGPT.

”We are going to see a neurologist for a new patient intake — what should we expect, what tests should we ask for?” she typed.

One suggestion was to talk to the doctor about doing an EEG, a test that measures electrical activity in the brain. In February, during Olivia’s first appointment with Dr. John McLaren, a neurologist at Children’s Hospital, Eaton brought up the test.

“EEGs are not necessarily done in kids that do not have clinical suspicion or clinical manifestations of seizure,” McLaren recalled explaining. But in Olivia’s case, they both agreed an EEG could still be worth trying. It would be a way of “sort of ‘peeking under the hood,’” McLaren said.

Less than an hour after getting the EEG, Olivia and her mother were shopping for fuzzy pajama pants — a reward for enduring the test — when Eaton’s phone rang. It was Dr. McLaren. He had the results.

“Hey, are you sitting down?” he asked.

“I was like, ‘No, I’m standing up in a fuzzy pajama pants shop,’” Eaton recalled responding.

The results showed Olivia had abnormal brain activity, which might indicate she was prone to having seizures. He was getting the pharmacy to fill a prescription for an anti-seizure medicine. Olivia should start taking it in the next hour.

Once on medication, Olivia finally began sleeping through the night. “This kid has not slept through the night since she was born,” Eaton said.

The journey continues

On a Saturday in August, Olivia, now 7, raced around her house in Winchester. Her father trailed behind her.

They needed to leave for Children’s, but Hospital Moose was missing. The stuffed animal with a blood pressure cuff on one leg wasn’t inside the duffel bag full of games, snacks, and other toys that Matt Eaton regularly lugged to the hospital. It was time for another monthly IV infusion of antibodies, a procedure that could take hours.

“Is this going to be her last one?” Fiona asked her mother.

It unfortunately wasn’t, Eaton responded. Although Olivia isn’t sick all the time anymore and gets enough sleep during the night, her health journey isn’t over. There’s no growing out of genetic mutations, but the Eatons said they feel a measure of relief knowing they’ve been able to get some answers and manage the symptoms.

They’ve also been diving deeper into Fiona’s genome. The bespectacled 9-year-old’s data revealed two mutations related to metabolic processing. From some initial research, Eaton learned that one of the mutations might be affecting how Fiona’s body processes creatine, which supplies energy to muscles.

Or, at least that’s what she thought for now.

“I’m not an expert yet,” Eaton said, “because I haven’t spent enough time with ChatGPT.”

Allyson Chiu can be reached at allyson.chiu@globe.com. Follow her on X @_allysonchiu.